技术规格

系统要求

不同量化方法和上下文大小的显存要求

DeepSeek-V3.2

DeepSeek-V3.2 is a powerful open-source Mixture-of-Experts (MoE) language model with 671B total parameters and 37B activated parameters per token. Built with an innovative architecture combining Multi-head Latent Attention (MLA) and DeepSeekMoE for efficient inference. Achieves exceptional performance across multiple benchmarks: 90.2% on MMLU-Pro, 84.5% on GPQA Diamond, 91.6% on MATH-500, 78.1% on Codeforces, and 92.3% on HumanEval. Supports 128k context window with strong multilingual capabilities. Features superior coding abilities, advanced mathematical reasoning, and competitive performance with leading closed-source models. Trained on 14.8 trillion diverse, high-quality tokens. MIT licensed for both research and commercial use. Ideal for complex reasoning, code generation, mathematical problem-solving, and general-purpose language understanding tasks.